For K-L10 (tracker inbuxa/inbuxa-charts#14), and the test half of K-C14 (charts #40).
What it does.INSTALL_WITH=argocd lab/e2e.sh (or flux) builds the lab as usual, then installs the chart through the controller instead of Helm:
Controllers: Argo CD v3.6.2 (core install) and Flux v2.10.1, each applied from its project's install manifest and checked by SHA-256 first (pins in versions.env). Flux publishes the sum; Argo CD doesn't, so its sum is the one taken when the pin was set.
Chart source: a Helm repository served inside the cluster (lab/gitops/gitops.sh repo). It holds the checkout's chart as 27.1.0-lab.1 and -lab.2, with the lab's values files packaged inside.
Secrets:
Under Argo CD: generateSecrets: false, and the webmail's and the metrics Secrets are made first.
Under Flux: the chart makes its own, as under Helm.
Self-heal is on in both: Argo CD's selfHeal, Flux's drift detection.
Then, after the usual client checks:
Steady. Three hard refreshes: in sync each time, nothing applied again, every Secret's data and uid unchanged.
Self-heal. The webmail scaled to 1 by hand comes back to 2.
Upgrade. -lab.1 to -lab.2 with every image pinned by digest, as a chart release pins them, while three webmail tabs stay open. The tabs must not be signed out.
After the upgrade. The same Secrets, mail through the load balancer, and steady again.
The 27.1 crossing from 2026.11.18.3 (chart/crossing-test.sh with GITOPS=), as changes to the controller's spec:
The controller is suspended while the old release sets up and fills the store.
Given rc.34 without the acknowledgement, the controller must give up, and the store is untouched.
Given it with the acknowledgement, the store is crossed and the mailbox digests match.
Results: all four gitops jobs pass (argocd and flux, amd64 and arm64, Kubernetes 1.37), in about 13–15 min each. The whole run is green: 13 jobs.
Install: 36 s under Flux, about 40 s under Argo CD after the data layer.
Upgrade: 32–42 s.
Crossing: about 190 s.
No Secret changed, and the tabs signed in once in every run.
Smaller changes:
fresh-store.sh now applies the spike's inbuxa-config itself. The crossing had only run on the bench, where it was left over from earlier.
Flux's upgrade remediation is off (retries: 0). A refused layout change has nothing to undo, and a rollback runs the same check (pre-rollback). That's for the GitOps docs page (K-D9).
Found: in 2 of the 4 upgrade runs, one webmail JMAP request (of 32–71) timed out after 30 s while all three components rolled at once. The webmail-only restart test (K-W5) has none. I'll find which roll causes it and file an issue separately; this PR only reports it.
For K-L10 (tracker inbuxa/inbuxa-charts#14), and the test half of K-C14 (charts #40).
**What it does.** `INSTALL_WITH=argocd lab/e2e.sh` (or `flux`) builds the lab as usual, then installs the chart through the controller instead of Helm:
- **Controllers:** Argo CD v3.6.2 (core install) and Flux v2.10.1, each applied from its project's install manifest and checked by SHA-256 first (pins in `versions.env`). Flux publishes the sum; Argo CD doesn't, so its sum is the one taken when the pin was set.
- **Chart source:** a Helm repository served inside the cluster (`lab/gitops/gitops.sh repo`). It holds the checkout's chart as 27.1.0-lab.1 and -lab.2, with the lab's values files packaged inside.
- **Secrets:**
- Under Argo CD: `generateSecrets: false`, and the webmail's and the metrics Secrets are made first.
- Under Flux: the chart makes its own, as under Helm.
- **Self-heal is on in both:** Argo CD's `selfHeal`, Flux's drift detection.
**Then, after the usual client checks:**
1. **Steady.** Three hard refreshes: in sync each time, nothing applied again, every Secret's data and uid unchanged.
2. **Self-heal.** The webmail scaled to 1 by hand comes back to 2.
3. **Upgrade.** -lab.1 to -lab.2 with every image pinned by digest, as a chart release pins them, while three webmail tabs stay open. The tabs must not be signed out.
4. **After the upgrade.** The same Secrets, mail through the load balancer, and steady again.
5. **The 27.1 crossing from 2026.11.18.3** (`chart/crossing-test.sh` with `GITOPS=`), as changes to the controller's spec:
- The controller is suspended while the old release sets up and fills the store.
- Given rc.34 without the acknowledgement, the controller must give up, and the store is untouched.
- Given it with the acknowledgement, the store is crossed and the mailbox digests match.
**Results:** all four `gitops` jobs pass (argocd and flux, amd64 and arm64, Kubernetes 1.37), in about 13–15 min each. The whole run is green: 13 jobs.
- Install: 36 s under Flux, about 40 s under Argo CD after the data layer.
- Upgrade: 32–42 s.
- Crossing: about 190 s.
- No Secret changed, and the tabs signed in once in every run.
**Smaller changes:**
- `fresh-store.sh` now applies the spike's `inbuxa-config` itself. The crossing had only run on the bench, where it was left over from earlier.
- Flux's upgrade remediation is off (`retries: 0`). A refused layout change has nothing to undo, and a rollback runs the same check (pre-rollback). That's for the GitOps docs page (K-D9).
**Found:** in 2 of the 4 upgrade runs, one webmail JMAP request (of 32–71) timed out after 30 s while all three components rolled at once. The webmail-only restart test (K-W5) has none. I'll find which roll causes it and file an issue separately; this PR only reports it.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
For K-L10 (tracker inbuxa/inbuxa-charts#14), and the test half of K-C14 (charts #40).
What it does.
INSTALL_WITH=argocd lab/e2e.sh(orflux) builds the lab as usual, then installs the chart through the controller instead of Helm:versions.env). Flux publishes the sum; Argo CD doesn't, so its sum is the one taken when the pin was set.lab/gitops/gitops.sh repo). It holds the checkout's chart as 27.1.0-lab.1 and -lab.2, with the lab's values files packaged inside.generateSecrets: false, and the webmail's and the metrics Secrets are made first.selfHeal, Flux's drift detection.Then, after the usual client checks:
chart/crossing-test.shwithGITOPS=), as changes to the controller's spec:Results: all four
gitopsjobs pass (argocd and flux, amd64 and arm64, Kubernetes 1.37), in about 13–15 min each. The whole run is green: 13 jobs.Smaller changes:
fresh-store.shnow applies the spike'sinbuxa-configitself. The crossing had only run on the bench, where it was left over from earlier.retries: 0). A refused layout change has nothing to undo, and a rollback runs the same check (pre-rollback). That's for the GitOps docs page (K-D9).Found: in 2 of the 4 upgrade runs, one webmail JMAP request (of 32–71) timed out after 30 s while all three components rolled at once. The webmail-only restart test (K-W5) has none. I'll find which roll causes it and file an issue separately; this PR only reports it.